{"id":"1a1e0311-d409-448d-85af-0403a5e61880","arxiv_id":"2607.02922","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SSM enrichment before hierarchical adaptive compression cuts video tokens 85% and yields 1.8× speedup while matching or exceeding uncompressed baselines on zero-shot reasoning segmentation.","lead":"STAC first enriches video tokens with linear-cost state-space scanning so redundancy can be measured, then hierarchically compresses them by ~85% for language-guided segmentation. The result is faster long-video reasoning masks that beat or match full-token baselines while supporting streaming.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The 85% reduction / surpass-baselines claim rests on a single-run, fixed-seed evaluation whose online–offline gap already shows the causal redundancy signal is incomplete for reasoning.","rationale":"The Reader correctly isolates the causal-SSM redundancy assumption and the online–offline gap as the soft spot; I agree that is the single most load-bearing concern. The paper’s architectural idea (enrich-before-compress with decoupled bi-spatial / causal-temporal Mamba) is cleanly ablated (Tabs. 3–4) and the zero-shot transfer story is interesting, so the contribution is real. What is missing for an unconditional accept is precisely the multi-seed, matched-budget evidence that would confirm the 85%/1.8×/surpassing numbers survive the incompleteness of the causal signal. That keeps the verdict CONDITIONAL, matching the Reader, with no need to escalate or downgrade.","tokens_in":18475,"tokens_out":564,"duration_ms":5381,"concrete_test":"Re-train STAC and the strongest full-token baseline (GLUS) three times with independent seeds under identical token budgets (STAC’s adaptive ~15% and a forced 15% GLUS ablation); report mean±std J&F on ReasonVOS online and offline. If the online STAC mean falls inside GLUS’s interval or the online–offline gap remains >2 points, the streaming-compatible “surpassing” claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (abstract, §4.2, Table 2) that STAC “achieves 85% token reduction and 1.8× speedup while surpassing compression-free baselines \to zero-shot streaming-compatible” is load-bearing on two untested premises: (1) that the cosine similarities of causal-SSM states (Eqs. 2–3) yield a task-aligned redundancy signal that adaptive thresholds (Eqs. 4–6) can safely discard without losing motion/reasoning tokens, and (2) that the reported J&F numbers are stable enough to declare “surpassing.” The paper itself records a 3.5-point online–offline drop on ReasonVOS (48.8 vs 52.3, §4.3), exactly the regime where future frames matter; yet every Table 2 entry is a single training run with no multi-seed variance, no confidence intervals, and no matched-token-budget re-run of the strongest full-token baseline (GLUS). Without those, the “surpassing under 85% compression” statement is not yet statistically or causally secured.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"STAC addresses the quadratic cost of dense video tokens in reasoning segmentation by placing selective state-space enrichment before hierarchical compression. Encoder features are first processed by a State-informed Spatiotemporal Aggregator (SSA) that applies bidirectional spatial Mamba scans within frames and causal temporal scans across frames (Eqs. 1–3), producing representations in which content redundancy is measurable. Hierarchical State-adaptive Compression (HSC) then performs temporal-then-spatial reduction with EMA-based adaptive thresholds (Eqs. 4–6), and the discrete retention decisions are trained end-to-end with a segmentation objective via straight-through estimation (Eqs. 7–8). Trained only on referring data (MeViS, Ref-YouTube-VOS), the method reports ~85% token reduction, ~1.8× speedup, and competitive or superior J&F on Ref-DAVIS17, MeViS, Ref-YouTube-VOS, ReasonVOS and ReVOS in a zero-shot, streaming-compatible setting (Table 2).","tokens_in":18805,"tokens_out":803,"duration_ms":7202,"significance":"If the results hold under multi-seed verification, the paper supplies a practical and architecturally clean solution to a genuine bottleneck: long-video reasoning segmentation under quadratic attention. The explicit decoupling of bidirectional spatial from causal temporal scanning is a useful design principle that simultaneously enables streaming and respects the distinct causal structure of the two axes. Task-grounded optimisation of compression thresholds via STE, the clean zero-shot isolation of reasoning benchmarks, and the public code release are concrete strengths. Ablations (Tables 3–4, Fig. 3) consistently favour the claimed design choices (temporal-first ordering, merge over prune, bidirectional spatial + causal temporal). The work is therefore of clear interest to the video-MLLM and efficient-vision communities.","major_comments":[{"comment":"The central claim that STAC “surpasses compression-free baselines … under 85% token reduction” (abstract, §4.2, Table 2) rests on single-run point estimates with no multi-seed statistics, error bars or confidence intervals. On ReasonVOS the reported margin over GLUS is only +2.5 J&F; the paper itself records a 3.5-point online–offline gap (48.8 vs 52.3, §4.3). Without variance estimates it is impossible to judge whether the “surpassing” statement is statistically secure.","section":null},{"comment":"Table 2 and §4.2 compare STAC (15% tokens) against full-token baselines (GLUS, VISA, VideoLISA, etc.) without a matched-token-budget re-run of the strongest full-token method. Consequently it remains unclear whether the gains arise from the SSM-derived redundancy signal (Eqs. 2–3, 4–6) or simply from any aggressive compression that forces the model to discard static background. A controlled ablation that forces GLUS (or an equivalent) to the same token budget would isolate the contribution of the proposed enrichment-before-compression pipeline.","section":null},{"comment":"The weakest modelling assumption—that cosine similarities of causal-SSM states constitute a reliable, task-aligned redundancy signal—is only partially stress-tested. The online–offline gap on ReasonVOS (§4.3) already shows that future-frame information matters for reasoning queries; yet the paper provides no quantitative analysis of which motion or multi-hop tokens are discarded by the adaptive thresholds, nor any failure-case study on long-horizon causal reasoning. Strengthening this analysis would make the claim that the recurrence-derived signal is “task-aligned” more convincing.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful idea here is simple and well executed: run a selective SSM to enrich tokens before you compress them, and do the scans asymmetrically—bidirectional spatial, strictly causal temporal—so the redundancy signal is already temporally informed and the whole thing stays streaming-compatible. That is a real step past VideoMamba/BIMBA/STORM-style unified bidirectional scans and past the usual “pool or prune the raw encoder features” baselines. They then fold the compression decisions into the segmentation loss via STE, train only on referring data, and show zero-shot transfer to ReasonVOS/ReVOS while claiming ~85% token cut and 1.8× speedup.\n\nWhat they do well: the ablations (Tables 3–4, Fig. 3) cleanly support the design choices—temporal-first ordering, merge over prune, bi-spatial + causal-temporal. The architecture diagram and equations are readable. Code is promised. On Ref-DAVIS17 and ReasonVOS they beat several full-token MLLM baselines; on MeViS/ReVOS they stay close while using far fewer tokens. That is a solid empirical package for an efficiency paper in this subfield.\n\nSoft spots, in proportion. Every main number is a single run with no seeds or error bars, so “surpassing” is not yet statistically locked. The paper itself reports a 3.5-point online–offline drop on ReasonVOS (48.8 vs 52.3), which shows the causal redundancy signal is incomplete when future frames matter—exactly the regime the abstract advertises as streaming-compatible. They also do not re-run the strongest full-token baseline (GLUS) under a matched token budget, so the efficiency–accuracy trade-off is a bit apples-to-oranges. Free parameters (EMA α, λ, the MLP for k) exist but are not the load-bearing problem; the missing variance and the online gap are. None of this sinks the contribution; it just means the headline claim is still provisional.\n\nThis is for people building long-video MLLMs or referring/reasoning segmentation who need linear-cost, online-friendly token reduction. It is not a foundational result, but it is a clear, citable engineering advance. I would send it to peer review; a serious referee can demand multi-seed numbers and a matched-budget baseline without killing the paper. Worth engaging if you work in this space.","headline":"Clean engineering paper: enrich with decoupled causal SSM first, then task-grounded hierarchical compression; 85% reduction with competitive zero-shot numbers, but single-run tables and an online–offline gap keep the headline claim provisional.","tokens_in":19422,"tokens_out":601,"would_cite":true,"duration_ms":6736,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"State-space recurrence first makes video token redundancy measurable, then lets a model throw away 85% of tokens and still beat uncompressed baselines on reasoning segmentation.","keywords":["video reasoning segmentation","token compression","state-space models","Mamba","spatiotemporal modeling","streaming video","multimodal LLMs"],"falsifier":"On ReasonVOS-style queries that require future context, measure whether online STAC (causal only) drops more than a few J&F points relative to an offline bidirectional version while still matching or beating full-token baselines; a large online gap that also undercuts the full-token score would falsify the claim that the causal redundancy signal is sufficient.","tokens_in":19398,"feed_emoji":"🎬","tokens_out":591,"duration_ms":40535,"temperature":0.7,"pith_summary":"Long video reasoning segmentation produces so many visual tokens that quadratic attention becomes impractical. Most compression schemes decide what to keep from raw encoder features that have never seen other frames, so they cannot tell semantic redundancy from surface similarity. The paper shows that a linear state-space recurrence can first condition every token on temporal context at linear cost; the resulting near-identical enriched states themselves become a reliable redundancy signal. STAC uses that signal: it enriches features with bidirectional spatial and causal temporal scans, then applies hierarchical adaptive compression whose thresholds are trained end-to-end for mask accuracy. The result is roughly 85 percent fewer tokens, 1.8 times faster inference, streaming compatibility, and higher scores than full-token baselines on both referring and zero-shot reasoning benchmarks.","feed_headline":"85% fewer video tokens, better reasoning masks","feed_subtitle":"State-space enrichment first, then adaptive compression, beats full-token baselines online","key_machinery":"State-informed Spatiotemporal Aggregator followed by Hierarchical State-adaptive Compression: bidirectional spatial Mamba scans plus causal temporal Mamba scans enrich tokens so that cosine similarity of the resulting states can drive adaptive temporal-then-spatial merging whose thresholds are optimised by segmentation loss via straight-through estimation.","core_discovery":"When encoder features are first passed through selective state-space recurrence, the recurrence itself produces a feature space in which content redundancy becomes directly measurable; hierarchical compression performed in that space can discard about 85 percent of visual tokens while still outperforming models that keep every token, and the causal half of the design supports online streaming.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["SSM first then cut 85% tokens, beats full-video baselines online","Recurrence makes redundancy measurable: 85% drop, stronger masks","STAC: hierarchical compress after state-space tops full-token seg","Causal-temporal SSM yields 85% fewer tokens with better results","Enrich then discard: 1.8× faster reasoning masks vs no compression"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That the near-identical states produced by causal recurrence are a trustworthy enough redundancy signal that adaptive thresholds can safely discard most tokens without losing the motion or future-frame cues needed for complex reasoning.","fun_headline_variants_meta":{"raw":{"variants":["SSM first then cut 85% tokens, beats full-video baselines online","Recurrence makes redundancy measurable: 85% drop, stronger masks","STAC: hierarchical compress after state-space tops full-token seg","Causal-temporal SSM yields 85% fewer tokens with better results","Enrich then discard: 1.8× faster reasoning masks vs no compression"]},"model":"grok-4.5","effort":"low","cost_usd":0.004298,"raw_usage":{"total_tokens":1277,"prompt_tokens":745,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":42980000,"prompt_tokens_details":{"text_tokens":745,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":452,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":745,"tokens_out":80,"duration_ms":4953,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:05:32.114986+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On ReasonVOS-style queries that require future context, measure whether online STAC (causal only) drops more than a few J&F points relative to an offline bidirectional version while still matching or beating full-token baselines; a large online gap that also undercuts the full-token score would falsify the claim that the causal redundancy signal is sufficient.","supporting_citations":[],"review_version":1}